From Robots to Animals:

Virtual Fences for Controlling Cattle

Zack Butler∗ Peter Corke† Ron Peterson‡ Daniela Rus§

Abstract

We consider the problem of monitoring and controlling the position of herd animals, and view animalsas agents with natural mobility but not strictly controllable. By exploiting knowledge of individual andherd behaviour we would like to apply a vast body of theory in robotics and motion planning to achievingthe constrained motion of a herd.

In this paper we describe the concept of a virtual fence which applies a stimulus to the animal asa function of its pose with respect to the fenceline. Multiple fence lines can define a region, and thefences can be static or dynamic. The fence algorithm is implemented by a small position-aware computerdevice worn by the animal, which we refer to as a Smart Collar. We describe a herd-animal simulator,the Smart Collar hardware and algorithms for tracking and controlling animals as well as the results ofon-farm experiments with up to 8 Smart Collars.

1 Introduction

Management of domestic animals is a very old human activity which has seen relatively little change overthe centuries. Cattle and other livestock forage over large paddocks whose perimeters are usually boundedby conventional fences. The two essential activities are confinement by fences and mustering or herding. Atypical ranch will normally have several paddocks separated by conventional fences. Animals are rotatedamong paddocks in an attempt to prevent overgrazing, and for husbandry practices such as during calvingor weaning. Fence technology has changed from stone and hedge to electrified wires, but is still labourintensive to install and maintain and is not easily reconfigured. Mustering has changed from shepherds onfoot perhaps aided by dogs, through the use of horses to motorized vehicles and even helicopters. Neverthelessherding remains a labor intensive activity when cattle and sheep graze over large paddocks and animals arerotated frequently to achieve pasture management constraints. It is physically hard work, often carried outin extreme weather conditions and remote locations. The costs of fencing and mustering remain a significantfraction of the cost of raising an animal, and has not benefited from the technical revolution in automation,computing and communication.

Within the robotics community there is much interest in human-robot interaction, but little on animal-robot interaction and this is surprising given the large number of domestic animals that we rely on, and theirimportance to our civilization.

Herds of animals such as cattle are complex systems. Herd societal behaviors are well known to stockfarmers but their study by animal ethologists [27] has just begun. There are interesting interactions be-tween individuals, such as friendship, kinship, group formation, leading and following. There are complexinteractions with the environment, such as looking for a water source in a new paddock by perimeter tracingalong the fence and random walking within the perimeter. Cattle have demonstrated spatial memory thatinfluences habitat use [14] and also memory capable of recognizing peers and humans [30]. Our goal is todevelop computational approaches for studying groups of agents with natural mobility and social interac-tions. Such systems differ in many ways from engineered mobile systems because their agents can move on

∗Computer Science Department, Rochester Institute of Technology, Rochester, NY 14623 USA, (e-mail: zjb@cs.rit.edu).†CSIRO ICT Centre, Australia, (e-mail: peter.corke@csiro.au).‡Dartmouth Computer Science Department, Hanover, NH 03755 USA, (e-mail: rapjr@cs.dartmouth.edu).§Computer Science and Artificial Intelligence Laboratory, MIT, Cambridge MA 02139, USA, (e-mail: rus@csail.mit.edu).

their own due to complex natural behaviors as well as under the control of the environment (for examplemoving toward a food or water source). We wish to generate models of such systems using observed physicaldata and to use these models to synthesize controllers for the movement of the mobile agents. Unlike morefamiliar robot control problems, the animal state (stress, hunger, desire) is only partially observable andonly limited control over motion can be exerted.

There are two fundamentally different approaches to controlling animal position: a physical agent suchas a sheepdog or robot, and a stimulation device worn by the animal. The first approach works by applying amoderate threat which the animal will flee from, however knowledge of the herd dynamics allows a very smallnumber of threat agents to control a very large number of animals. In this category there is the pioneeringwork of Vaughan [29] who demonstrated a mobile robot that was able to herd a flock of ducks to a desiredlocation within a circular pen.

The second category relies on an active device worn by the animal that can provide an aversive stimulusas required, typically this is an odorous spray, a sound, or an electric shock. Devices to control domesticpets such as dogs are commercially available, either to prevent it from barking or to constrain it. The latterdevice typically works by sensing proximity to a buried perimeter wire which allows for a simple collar onthe dog. Clearly such an approach, based on installed infrastructure, would be prohibitively expensive forlarge scale agriculture and has all the disadvantages of a conventional fence (installation, maintenance andnot-configurable). However with the advent of GPS technology, and the rapid reduction in receiver cost, itis possible to implement such a fence without installed infrastructure. That is, a boundary can be definedin terms of geographic coordinates and a stimulus applied when the animal moves beyond the boundary.This functionality replaces that of a fence and is referred to as a “virtual fence”. By making the boundarydynamic we achieve the function of herding or mustering, almost for free.

In this paper we describe our work which brings together concepts from the animal behaviour andmanagement communities, adhoc networking and robotic motion planning. In [9] we described our firstexperiments in controlling a herd of cows with a single static virtual fence using an approach that relies onad-hoc networking. In [8] we extended that work on the algorithmic side, by introducing motion planning forcomputing dynamic virtual fences whose goal is to muster the herd to a new location. We have implementedthese algorithms in simulation and deployed 10 smart collars on cows at Cobb Hill Farm in Vermont.

Our work builds on important recent studies in studying and modeling animal behavior. The Leurreproject [1,10,11,13,15,17,19,21] has contributed seminal theoretical modeling methodologies for individualsin animal societies, based on differential equations whose parameters are computed from exponential distri-bution of times in animal states from real data. This approach seems to model well ants, cockroaches, andfish. However, modeling the interaction between individuals and modeling large animals remains challenging.Vaughan [29] demonstrated a mobile robot that was able to herd a flock of ducks, and Denebourg etal [11]were able to influence the behaviour of cockroaches by the motion of a small robot. Within the networkingcommunity there is growing interest in using GPS technology to track animals with fine resolution. TheZebranet project [16] considers the tracking and data collection issues in tracking herds of zebras in theirnatural habitat. Related work on modeling and coordination in animal and artificial robot societies alsoincludes [4–7,20, 22, 24].

This paper is organized as follows. Section 2 describes the prior art regarding virtual fences, our al-gorithmic implementation, a detailed simulation model, and the extension to dynamic path planning. InSection 3 we describe the experimental approach in some detail including the hardware and software toolswe developed to facilitate the experiments. Section 4 describes experimental results that cover: collectingdata to create a grazing model for the cows; collecting connectivity data and information propagation datafor the multi-hop routing method within the herd; collecting stimulus-response data for individual animals;and collecting response data for the virtual fence on a group of animals.

2 Virtual Fencing

The application of smart collars to control cattle is first discussed in detail by Tiedemann and Quigley [23,28]who were concerned with controlling cattle grazing in fragile riparian environments. Their first work [23],published in 1990, describes experiments in which cattle could be kept out of a region by remote manuallyapplied audible and electrical stimulation. They note that cattle soon learn the association and keep out

����

����

����������������������������������������������������������������������������������������������������������������������������������

��������������������������������������������������������������������������������������������������������������

�������������������������������������������������

�������������������������������������������������

virtual fence

cow

�

�

�

���

Figure 1: Notation used for the virtual fence algorithms.

of the area, though sometimes cattle may go the wrong way. Cattle learn to associate the audible stimuluswith the electrical one and they speculate that the acoustic one may be sufficient after training. Morecomprehensive field testing in 1992 is described in [28] and also identified issues about the need to trainanimals to associate stimulus with spatial restrictions. This issue is also discussed by Anderson [2].

The idea of using GPS to automate the generation of stimuli was proposed by Marsh [18]. GPS technologyis widely used for monitoring position of wildlife. Anderson [2,3,12] builds on the work of Marsh to includebilateral stimulation, different audible stimuli for each ear so that the animal can be better controlled. Theactual stimulus applied consists of audible tones followed by electric shocks.

The choice of stimulus is an important component of working with animals. The animal behavior lit-erature discusses both sound and electric shock, as well as their combination. If a sound stimulus alwaysprecedes an electric shock then the animal will become trained over time and react to the sound rather thanwaiting for the shock. If sound is used this raises the question about the optimum sound to use. Should thesounds be natural or unnatural, can they encode information about the direction to turn, and so on.

In our experiments we used sound stimulus only for several reasons. First was an admitted squeamishnesson our part in shocking cattle — the cows we worked with are a small herd of dairy cattle owned by acooperative and are very much like pets, with individual names such as Linden and Hazel. Second, theexperimental animal protocol approved by Dartmouth College allowed only the use of gentle stimuli suchas sounds (protocol assurance A3259-01). Finally, we wished to explore and characterize the scope of aminimalist and friendly approach to controlling cattle that relied on sound only.

2.1 The Virtual Fence Algorithm

In our algorithms, a fence is represented by a point Fp and a normal vector Fn. The notation used is shownin Figure 1. The point Fp is given in (latitude,longitude) coordinates, although higher-level interfaces such asthe GUIs described in Sec. 3.3.3 use relative distances internally and convert to the absolute coordinates whensending the fences to the collars. Fp can be simply any point along the fence, and Fn points perpendicularlyinto the interior half-plane.

This representation allows a simple calculation to determine whether the cow is “outside” the fence.The collar first computes the cow’s position relative to the fence xr by subtracting the Fp from the cow’sabsolute position xa. The distance d that the cow is behind (or in front of) the fence is then computedas the dot product of xr with Fn. If d is positive, the cow is in the desired region, while if it is negative,the cow is behind the fence. To compute the exact distance correctly, xr is first converted to meters, sincethe Fn is defined based on equal units in each direction. This is all done with nothing more complex thanmultiplication.

For most applications, several fences will be present. Since each gives an interior half-plane, setting up anumber of fences with their normals pointing toward each other results in a convex polygonal paddock. Todo position checking in this case, the algorithm computes a d value for each fence, and reports the fence anddistance that the cow is farthest behind (the most negative d). This is necessary to ensure the cow does notappear to be close to the interior when it is in fact far behind another fence. The algorithm, with stimulustimeout, is given in Algorithm 1.

Algorithm 1 Virtual fence. Tmax is the maximum number of sequential stimuli that can be applied andTtimeout is the inhibitory period that applies after that limit is exceeded.

1: tstimulus ← 02: ttimeout ← 03: loop

4: dmin ← 05: for i = 0 to Nfence do

6: d← (xa − Fp) • Fn

7: dmin ← min dmin, d

8: ttimeout ← ttimeout − 19: if d < 0 ∧ ttimeout < 0 then

10: Cow ← stimulus11: tstimulus ← tstimulus + 112: if tstimulus > Tmax then

13: ttimeout ← Ttimeout14: else

15: tstimulus ← 0

To present a stimulus, the simplest option is to produce a sound of a given volume when the cow goesbehind the fence. This can be done by testing the sign of the cow’s fence distance. However this is quitean uninformative stimulus and the animal can only determine the desired position by moving about untilthe stimulus stops. Instead we encode the distance from the fence to provide graduated stimulus whichgets stronger the farther the cow is behind the fence. This can be done by generating a sound with volumeproportional to −d. Finally, a more complex procedure is to monitor d over time, and stop the stimulus assoon as the cow begins to move toward the desired region. This is done by keeping track of dmin. If thecow’s current distance is greater, a stimulus is not produced. A further sophistication would be to reducethe stimulus when the animal turns in the right direction, providing a reward for the correct response. Inthe approach of Anderson [3] a directional stimulus is applied by left and right sound sources as well aselectrodes on each side of the neck.

If the fence is to move, it is instantiated with a non-zero velocity Fv, in m/s. The point is then movedas a function of time along the normal

Fp(t) = Fp(0) + γFnFvt

where γ is a scale factor from northing/eastings to lat/long coordinates. The fence orientation could also bea function of time, that is Fn(t).

Figure 2: Screenshot from the simulation of the virtual fence algorithm. Twenty cows are represented byellipses, and one fence is shown as a vertical line.

2.2 Modeling and Simulation

To test our ideas we developed a simulator that models the behavior of a herd of cows. We were inspired byVaughan’s duck simulator [29], but extended the animal model to account for the differences between thespecies as well as their environments, and added the capability for virtual as well as physical fences, animalpaths, and even wind.

2.2.1 Animal model

To model the behavior of a cow, we use potential fields to capture response to external stimulus along withinternal state (which also affects motion). It is well known that a cow’s movement will be dependent on thelocation of its herd. Therefore, we model the effects of one animal’s position on another’s motion: a weakattractive herding field, and a strong local “personal space” field. We also use potential fields to model theresponse to non-cows in the environment, such as a robot, human, or other obstacle, which represent a threatand generate a strong repulsion. A fixed fence exerts a strong normal repulsive force over a short range,whereas the virtual fences are modeled through a different method, since they do not have remote effect.

From an internal point of view, we simplify the animals’ desires with a two-state behavior model, walkingand grazing, each with an associated speed and duration distribution. Transitions between these statesoccur in a stochastic fashion and are used to induce motion in an approximately forward direction. We alsoexplicitly model the stress of each animal and use this to affect the animal’s behavior. Stress is created bythe virtual fence stimulus as well as the nearby presence of other fast-moving animals, isolation from theherd or threats. An animal in a low stress condition will alternate between grazing and walking as describedabove and exhibit very little herding instinct (as observed in the field) unless it becomes well separated. Ananimal that is experiencing high stress will move toward other animals, and will not resume grazing until itsstress has gone down. The stress level of an animal decays exponentially over time.

In addition to inducing stress, the fence stimulus has an immediate effect on the motion of the animal.We have used two different models, each of which take inspiration from field observations. In the first model,a stimulus causes the animal to quickly turn approximately 90◦. This behavior was also observed in [23]. Inthe second model, the cow walks forward for a short time when stimulated.

2.2.2 Simulator details

We have implemented the animal model described above within a Matlab simulation that also includes staticand virtual fences. The simulator displays the position and orientation of each animal as well as a graphical

(a) (b)

Figure 3: Screen shots from a simulation of moving virtual fences, 20 cows represented by ellipses in a50m× 50m space, with four virtual fences forming a moving rectangle around them. (a): the beginning ofa simulation. (b): after 4000 time-steps which correspond to 45 mins. See also Extension 4.

indication of the animal’s stress level. The simulator instantiates a number of animals in a moderately closegrouping, giving each animal low stress and a random internal state. This also ensures that the simulatorstarts in a stable condition. The simulation then enters a loop in which time is advanced by small amountsand each cow’s position, velocity and stress is updated. Figure 2 shows a typical simulator display.

At each time step, the simulator determines each cow’s position relative to the rest of the herd andcomputes a potential-field-based force from these relative positions. If the animal is distant from all others,we add an additional large force toward the herd and increase the its stress level. Fast moving or stressedanimals in close proximity add to the current animal’s stress. The simulator then checks the cow’s positionrelative to all fences, real and virtual. For the real fences, if the animal is in close range, an additional forcenormal to the fence is added to the force acting on the animal. The virtual fences induce a stimulus whenthe cow passes into the fence zone, and the impact of this stimulus is to provide a large force as describedabove. We assume that this stimulus will be strong enough to override any other behavior. Thus, for thetime step when the stimulus is present, it alone is used to determine the cow’s motion. The simulationspresented here are based on a bilateral stimulus, such that the cow will turn sharply away from the fence(but not based on its exact angle with respect to the fence).

After the total force acting on the cow is computed, we consider the cow’s internal state to determine itsmotion. At this time the simulator also determines if it is time to switch to the other state. During a stateswitch, the animal can make small heading changes and the simulator stochastically chooses the amount oftime to spend in that state (barring external influences). If the induced force is small, this is ignored andinstead a small forward force is used to generate a speed appropriate to walking or grazing.

To determine the influence of the induced force on the cow, we model the cows as non-holonomic andgive them a maximum angular velocity. If the virtual force given by the potential fields is not closely alignedwith the cow’s current direction, the cow will turn to align itself with the force. Otherwise, the force willaffect the cow’s velocity, and in either case, the position will be updated based on the velocity and the sizeof the time step.

2.2.3 Simulation results

We have performed a variety of simulations with different numbers of animals, different parameters for thepotential field calculations, and different fence configurations. Multimedia extensions 1–3 show unforcedbehaviour, and behaviour with slow and fast moving threat objects. In one class of experiments, presentedas Fig. 3 and Extension 4, we started with a group of 20 cows within an overall 50×50 m area and put virtualfences in a rectangle closely surrounding them. We then allowed the interior area to move without changingsize, and the animals did move accordingly. We have also experimented with selective virtual fences, that is,fences that affect only certain of the animals in the herd. In simulation, this allows the cows to be separatedinto two groups, as seen in Fig. 4 and Extension 5. Lines drawn from a cow indicate a force vector which isthe result of a virtual fence interaction.

We note here that the effectiveness of all the mustering experiments does depend on the potential field

(a) (b)

Figure 4: Simulation of selective mustering. (a): at the beginning of the simulation, the two classes ofanimals, represented here by dark and light ellipses, are intermixed. (b): after a short time, the two groupshave been separated. See also Extension 5.

parameters and stimulus response used. We have seen good results with a variety of internal parametersprovided that the stimulus is moderately effective. Our best results were obtained for the simulation of abilateral stimulus, under which the angle between the cow and the fence is calculated and a stimulus providedthat moves the cow a constant amount away from the fence. We have also tried varying the likelihood thatan animal will respond to a given stimulus, and found that often 100% stimulus is not required, but that ifan animal gets far beyond the initial fence location it may not be able to return to the interior.

To test the algorithms against these models, we ran virtual fences on a simulated herd with widely varyingparameters. The overall goal was to move the virtual fence slowly into the herd and test how quickly theherd moved away from the encroaching fence. This was tested with different values for the grazing speedand walking speed of the cows, the level of herd-attraction and the probability that a stimulus would havethe desired effect. We found that the parameters affected the overall speed of the herd in front of the fenceand the number of stimuli that were applied, but in all cases the herd did move in the desired direction.

We also tested the effect of using cow orientation in controlling stimulus. Our expectation was that ifthe cow tends to go forward when stimulated, it would be necessary to sense the cow’s orientation and onlyapply stimulus when the cow is pointing in the direction we wish it to move. In the simulation, this turnedout not to be necessary, since after receiving the stimulus, the cow would have increased stress and returnback toward the herd even if it initially went the wrong way. However, this behavior is very dependent onthe nature of the stress model.

2.3 Planning for Dynamic Fences

An important goal of this work is to automatically muster cows from one pasture to another. To make thispossible, we have developed a path planning system for virtual fences. While this problem shares some basicfeatures with traditional robot path planning, it has important and interesting differences as well.

The planner creates a path from one point to another using a simple occupancy grid and A* search [25].The object executing the plan is a virtual (polygonal) paddock, with significant extent that may change asit moves1, as long as its area remains sufficiently large for the herd. Thus, it is easier to perform planning inthe workspace than the configuration space. Obstacles can overlap somewhat with the virtual fence edges,changing the effective area of the virtual paddock but not altering the plan. We use planning operatorsthat change the dimensions of the paddock while keeping the amount of free space within the paddocksufficient for the given number of animals. Finally, we would like the motion to consist of a small numberof straight-line segments. This type of optimization is necessary because changing the animals’ direction ismore difficult and confusing than keeping them moving along their current vector, and to limit the numberof fences that are downloaded to the collars. This optimization can be implemented by giving turns a largecost in the A* search.

To create a plan, instead of doing an expensive search in five dimensions (paddock location, width, height

1Both extent and shape may change.

(a) (b)

Figure 5: GUI for dynamic fence planning. Environment is based on the Dartmouth Green, with somereal-world paths defined to be obstacles. In (a), finding a path from the start (in top portion of free space)to the goal (at lower left) requires a significant change in the size and shape of the virtual paddock. (b) Amore complex path — note that small overlap with obstacles is considered acceptable as long as sufficientfree space remains in the virtual paddock.

and motion direction) we first generate a pair of baseline plans. The first uses a point-sized virtual paddockand the second a constant-sized square paddock. These plans can be computed using the motion operatorsonly. If the two plans are similar in length and complexity then the square paddock plan can be used asis. If not, we then gradually reduce the size of the square paddock until a plan is successful. The successfulplan is then used as a base for a path with a variable-sized virtual paddock. The search for this final path isdone efficiently in three dimensions (width, height and distance along the base path.) The GUI developedfor this planner along with examples of resulting plans is shown in Fig. 5.

Plans are turned into schedules of fences that are executable by the collar. For each segment of the path,four fences are required to define the paddock. A velocity is set for the fences, which in turn gives eachsegment a time interval over which its fences are active. This list of fences, with a point, normal vector, speedand relative time interval for each, is then given to the collars. An example of this type of fence schedule isgiven in Sec. 4.5. The collars wait for an initialization message which tells them to make the time intervalsabsolute from that moment and continue to evaluate the fences and make appropriate stimuli for the courseof the path.

It should be noted that the planner is only planning for the virtual fences, and assumes the desiredbehavior of the animals within. In the future, we plan to validate quantitative models for the animals’reaction to the stimulus. This will tell us how effective the moving fences will be, and in turn how the fencesshould be moved to produce the desired motion of the herd. It may be possible to make the motion adaptiveto the position of the animals and only move when all animals are away from the advancing fenceline.

3 Experimental setup

3.1 The farm

Cobb Hill farm is a cooperative dairy farm in Vermont. An aerial view is shown in Figure 6. The three fieldsin which experiments were conducted are shown outlined in black. Field 1 is a long narrow field on the sideof a steep hill, the top of which is on the left in the photo. There are two strips of bushes and trees dividingthe field into three parts. The cows can walk around the dividers to reach the higher pastures. The treesin the dividers and the trees around the edge of the field provide shade. Field 2 is a larger field with muchmore open area. It is also on the side of a hill, the top of which is the left edge of the field in the photo.Trees around the West and South sides provide shade. The field is open with few obstructions (a stand oftrees near the north end and a few steep inclines.) Field 3 is a narrow pasture that gradually widens out.

Figure 6: Aerial view of Cobb Hill farm. The fields where experiments were conducted are outlined in black.North is up. The photo displays an area approximately 1 km on a side.

Trees on the east edge provide shade. The land in field 3 is relatively flat with no obstructions to movement.A water trough is located half way up field 1 on the South side, and at the entrances to fields 2 and 3.

Collars were fitted to the cows in the barn in the morning and they were then released for the day. Theirhabit is to move fairly quickly to the far end of each field, and then wander around slowly till evening. Theexperiments were conducted in summer and fall of 2003.

3.2 The Smart Collar Hardware

Prototypes of a Smart Collar were constructed using commercial off-the-shelf components that are readilyavailable. Figure 7(a) shows the components of a collar. The computer is a Zaurus PDA with a 206MHzIntel StrongArm processor, 64MB of RAM, with an additional 128MB SD memory card. It runs EmbedixLinux with the Qtopia window manager. The Zaurus has a serial port and stereo sound port. A Socketbrand 802.11 compact flash card provides a wireless network connection. An eTrex GPS unit is connectedto the serial port of the Zaurus. A small Smokey brand guitar amplifier is used to reproduce sounds fromthe Zaurus audio port. A fully assembled collar is shown in Figure 7(b). Figure 7(c) shows a cow wearingan early version of the collar.

The counterweight is required to stop the collar from rotating, the GPS receiver and PDA WiFi cardneed to be on the top of the neck to ensure good satellite and radio reception. However in practice the collarsdo rotate somewhat. A very practical problem is water from long grass and early morning dew, or whenthe animal drinks at the trough and the lowest part of the collar goes in the water. The collar is not fullywaterproof, though it is fairly water resistant since the GPS and audio amp are well sealed and the speakerhas a plastic cone. The Zaurus is enclosed in a plastic case which gives it some water resistance, althoughthe holes for the cables will allow some water in if the collar has rotated. The batteries in the Zaurus are thelimiting factor in how long the collar will run, giving about two hours and forty minutes of life. The audioamplifier and speaker will produce about 90 to 100dB volume at a one foot range, depending on the natureof the sound.

The Zaurus has a custom kernel which allows running the WiFi card in ad-hoc peer-to-peer mode. Thisallows us to do multihop forwarding of messages for better connectivity within the herd. Ssh and scp areinstalled to allow remote login to the Zaurus and field upgrades of software. Each Zaurus is also configuredwith a shell terminal program and has a foldup keyboard for accessing and running programs directly fromthe console.

A laptop computer is used as a basestation for sending commands to the collars. A Cantenna branddirectional WiFi antenna is used with the basestation to improve communication range to the herd.

and

CounterweightCounterweight

Audio Cable

���������������������������������������������������������������������������������������������������

���������������������������������������������������������������������������������

GPSZaurus

WiFi Card

Speaker

Amp

Nylon Collar

Top

Serial Cable

(a) (b)

(c)

Figure 7: (a) The components of the Smart Collar include a Zaurus PDA, WiFi compact flash card, eTrexGPS, protective case for the Zaurus, an audio amplifier with speaker, and various connecting cables. (b) Afully assembled Smart Collar, with PDA case open. (c) A cow with a collar in Field 3.

Figure 8: Peter Corke and Zack Butler log into the networked cows to download code during an experimentat Cobb Hill Farms.

Figure 9: Volume of sound produced for Zaurus volume settings. The ”C Weighted” notation indicates thesound level meter has applied a filter to adjust for the frequency response of the human ear. The A Weightedcurve does not have this compensation.

3.3 Software Infrastructure

The components of the software used in the experiments are as follows:

3.3.1 Virtual Fences

A fence is essentially defined as a point on the surface of the earth, an “interior” direction, and a velocity.The velocity can be used to move the fence over time toward or away from the specified direction. Fencescan be added or removed at any time and several of them can be created at once from definitions stored in afile. Several fences can be combined to create convex polygonal shapes. When the GPS readings indicate acow has crossed a fence a sound is triggered. The sounds are stored in WAV format files and can be selectedfrom a list to be played on the Zaurus audio device. The sounds used in our experiments were natural suchas cows and dogs; more alarming such as lions, tigers, panthers, wolves and thunder; and unnatural such asa car crash and a helicopter2.

The volume of sounds is controllable on a percentage scale from zero to 100 percent. All fences use thecurrently selected sound and volume, which can be changed without redefining the fences. Figure 9 showsthe relationship between the Zaurus sound settings, which are expressed on a percentage scale, and theactual volume produced by the Zaurus/amplifier combination. The curve becomes flat around 40 percentdue to the amplifier becoming saturated and starting to clip. This results in a progressively harsher soundas the volume is raised which is perceived as louder, but which is not actually louder.

An early version of the collar was built with two speakers to allow evaluation of stimuli with a directionaleffect by differential volume control. However we were unable to control the “balance” on the Zaurus soundcard so we reverted to a single speaker.

The fence module also reads and interprets the GPS data which arrives every two seconds when the GPShas a good lock on the satellites. It also sends a periodic Alive message indicating the collar is functional.

3.3.2 Message Handling

Wireless network and Unix pipe messages are used to control the software. The same message format is usedinterchangeably for both message types. This allows messages to be sent locally, which is useful for testing,and remotely via WiFi for field experiments. All WiFi messages are multihop, being forwarded once by eachcollar, to improve range and connectivity within the herd. There are two message channels, one outgoing

2Our experiments were done under protocol assurance A3259-01 given by the Institutional Animal Care and Use Committee(IACUC) of Dartmouth College.

Pipe

WiFi Multihop

Pipe

Pipe

Pipe

Sender

Receiver

Receiver

Sender

Control

CommandsSent

RepliesReceived

Acks &Alive

Messages

Pipe to Pipe for Testing

Pipe to Pipe for Testing

WiFi Multihop

Collar

Listen

Figure 10: The message routing structure for the networked cattle herd. Messages are routed along multi-hoppaths from animal to animal in the herd, ie. a simple flooding protocol. There are two message channels,once outgoing from a basestation and one incoming to the basestation.

from a basestation and one incoming to the basestation. The outgoing channel is used for defining fences,manually triggering sounds, setting sound type and volume. The incoming channel carries “Alive” messagesindicating a collar is active, and acknowledgment messages for receipt and proper interpretation of messages.

3.3.3 Experiment Control

Both text and GUI control programs are used to manage the collars in the field. The text control program canbe run on a Zaurus or Linux laptop and allows setting and deleting fences, setting type and volume of sound,and manually triggering a sound. The GUI control programs, written in Tcl/Tk, include the functionalityof the text program, and provides buttons for triggering sounds on specific cows, a map display showingcurrent cow locations and status (i.e., relationship to fence boundary and whether a sound is playing), anda status display showing whether Alive messages have been received recently from each cow.

Figure 11 shows the control GUIs. The software programs on the collars and basestation are started andstopped with shell scripts for easy reconfiguration.

3.3.4 Logging and Time Synchronization

A variety of information is logged on the collars for experimental and debugging purposes including GPSlocation, GPS time, messages received, messages forwarded, and messages sent. All log entries are accom-panied by a time and date stamp. To ensure accurate timestamps across the several programs in the collarsystem the Zaurus clock is initially synced to the GPS timestamp. Then a gettimeofday() system call isused in the various programs to log the current time. The drift in the Zaurus clocks is sufficiently low toprovide good time sync for the duration of experiments which typically last two or three hours. Log data ispost-processed using custom written scripts in a variety of languages.

3.4 Experimental Methodology

We used the methods of direct observation, video taping, and note-taking to evaluate the effectiveness ofthese sounds in producing desirable reactions. We used the GPS measurements of position and velocity tostudy the cows’ reaction to sounds. Did they avoid spaces beyond fences? Did they change direction? Didthey change walking speed? Looking for correlations between sound events and changes in the GPS data wasa primary analysis method, though we found it limited by the spatial accuracy of the GPS position data.

4 Experimental results

Our experiments addressed four issues: (1) collecting data to create a grazing model for the cows, whichis used in the simulation and fence control algorithm; (2) collecting connectivity data and informationpropagation data, which is used to assess the performance of multi-hop routing in this environment; (3)

(a) (b)

Figure 11: GUIs used on laptops to monitor field experiments. (a) Sound control GUI. Pressing a buttontriggers the current sound on a specific cow. Current sound and volume can also be selected. (b) Map controlGUI. Shows the last reported position of each cow, whether it is currently playing a sound, and whetheran Alive message has been received recently. Buttons and text boxes in upper right show recent commandacknowledgments from collars.

collecting stimulus-response data for individual animals; and (4) collecting response data for the virtualfence on a group of animals.

Our preliminary results are encouraging. Animals respond to artificial potentials of sounds generatedby the virtual fence by moving forward if they are on their own, or toward the group if they are in closeproximity to the group. The animals responded to sounds (see Figure 13(b)) but habituation to stimuli wasa problem. Others [3] have combined the sounds with shocks to avoid habituation.

4.1 Acquiring a Grazing Model

Methodology Our first field experiments were conducted to attempt to verify the two-state grazing modelused in simulation. The first experiment involved five collars placed on cows which were released into Field 1.These collars were populated only with the GPS devices and used their built-in tracking function. However,this function is designed to track human hikers who tend to move at a higher and more constant speed, andso this did not give sufficient temporal resolution to test our models. A second experiment with eight fullcollars on cows in Field 2 allowed for better collection of data, which we were then able to analyze. Theeight collars were put on the cows in the morning, and the PDAs recorded GPS positions every two secondsuntil the battery ran out.

Data Figure 12 shows a GPS location track for eight cows moving as part of a herd of 14 cows over time.They start out in the barn, follow a path to the field, and then wander to the far side of the field and back.In addition to the position tracks, the timestamps on the graph give a rough idea of how the herd movedand how spread out they were over time. For an idea of scale, the trek from one end of the field to the othercovered a distance of about 300 meters.

In order to look at an appropriate sample of the cows behavior, we present only the data from after thecows had reached the field. A histogram of the velocity for one cow is presented in Fig. 13. Each samplerepresents the difference in consecutively recorded positions, usually two seconds apart. Due to the resolution

Figure 12: GPS position track of eight cows, moving as a herd, over time. A few timestamps give some senseof how the cows wandered with respect to each other.

0 0.5 1 1.50

50

100

150

200

250

300

350

400

450

Speed [m/s]

Num

ber

of o

ccur

renc

es

0 0.5 1 1.50

50

100

150

Speed [m/s]

Num

ber

of o

ccur

renc

es

(a) (b)

Figure 13: Histogram of speed of one cow over a period of 40 minutes. (a) Based on raw GPS differences(b) Based on a 10-second moving average speed. Note difference in vertical scales.

(a) (b)

Figure 14: (a) Time history of Alive messages received by collar 9 in a typical experiment. (b) Averagenumber of hops for an Alive message to reach the basestation.

of the GPS data, we also present a 10-second moving average of the speed data. These plots represent datacollected from one cow, but other cows in the herd display very similar overall velocity profiles.

Discussion These data show that the cows have a range of speeds throughout the day, and the distributionis approximately bimodal: we see that the animals spend a large amount of their time moving quite slowly,and the rest of the time at higher speeds. The higher speeds vary. During our field observations we notedhigher speeds when an animal moved from a grazing area to another. Even higher speeds were observedoccasionally when it tried to catch up when disconnected from the rest of the group and when the animalseemed to respond to an external fearful factor such as a human or a barking dog. The speeds in the firstcategory seemed to depend on the distance gap between the animal and the rest of the herd. The averagespeed for the grazing behavior is under 0.2 m/s. This can detected reliably by consecutive GPS readings withdifferential GPS, and it also can be detected with lower-end GPS (like the one we used) from the smoothedspeed data. For Cow No. 10, using the smoothed speed gives a smooth distribution that peaks at around0.16 m/s. Setting a cutoff for the grazing of 0.4 m/s gives a mean grazing speed of 0.167 m/s. The cows alsospent a significant time at higher speed, walking from one grazing spot to another. This behavior takes placeabout 15% of the time (183 raw samples or 165 smoothed samples above 0.4 m/s from 1200 total samples).However, the walking takes place at a uniform range of speeds up to 1.25 m/s, rather than a single walkingspeed as originally supposed.

4.2 Network connectivity

Methodology The collars periodically emit an “Alive” message which are propagated by the multi-hopmessage layer using a form of flood forwarding. Each collar also logs all “Alive” messages that it receives,along with the sender id and time.

Data Figure 14 shows an example of the connectivity achieved between collars over time. In the first halfof this graph there is very good connectivity during the time the cows were all together in the barn. Around30,000 seconds most connectivity is lost as the cows are walking end to end along a narrow path out to thefield. On the right side of the graph connectivity varies as the cows wander around the field, their bodiesand tall wet grass being the main causes of signal obstruction. Connectivity to collar 7 is lost before 32000seconds because of an equipment malfunction. Collar 6 was not deployed. Each collar was connected to atleast one other collar 97.5% of the time. On average, each collar was connected to 66% of the rest of thecollars, via one or more hops. Thus there is room for improvement in the multihop routing algorithm whichis based on simple forward-once flooding.

Figure 14 shows the number of hops required for an Alive message to reach the laptop basestationduring an experiment. Most messages are relayed only once to reach their destination which indicates good

connectivity between collars.

Discussion Dynamic graphs of the message routing have shown us that connectivity among the herd isusually quite good since the cows tend to stay near each other. Connectivity with the base station wasproblematic in that there is a trade off in staying far enough away to not influence the herd (they are verycurious and friendly) and staying close enough to maintain radio contact. WiFi networks are essentially lineof sight and are blocked completely at times by the cows bodies. Switching to VHF transmitters to improvebasestation connectivity is an option we are considering.

4.3 Individual Response

Methodology In all field experiments we visually observed the behavior of individual cows, both awayfrom the herd and when amongst the herd. We often video taped moments of interest, see Extension 6, andhave the data logs from all experiments as a basis for analyzing individual behavior. GPS velocity data wasalso used to correlate stimulus events with changes in velocity.

Data A single sound sometimes had an effect on some of the cows and had no effect when tried on othercows. Some cows never reacted to sounds, while others were more sensitive. Observed reactions to a stimulussound varied widely and included stop eating and look up, no reaction, stop eating, look up, then walk ashort distance, usually forward, etc. (see Table 1).

The velocity of an individual can be derived from the GPS tracking data to discover if a cow reactedto a sound stimulus by changing speed. Figure 15 shows a time history of the speed for a cow in oursecond experiment. The asterisks denote when a sound was played. In this case there seems to be a goodcorrelation between sound events and the cow being in motion. However, some of the animals respondedin a less correlated way. Two difficulties in interpreting this kind of data are, first, cows may already bein motion when stimulated, and second, the GPS data is very coarse in time (a reading every two seconds)which makes it difficult to judge if the cows motion was actually in reaction to the stimulus.

Table 1 shows some of the observed responses. We generally noted that repeated and louder sounds weremore effective in eliciting a response. Cows would often react to the first instance of a sound and then notreact to further instances. Waiting a half hour would sometimes result in them reacting again to an initialsound.

Discussion Some cows definitely reacted strongly to a sound stimulus, though they often quickly becameinured to it, and stopped reacting. The orientation of the cow before the stimulus was applied played a rolein determining what direction a cow moved, if it moved. Further research into effective stimulus methodsand into invoking directional behavior are needed. Much louder sounds may be more effective. Soundsaccompanied by something visible such as a puff of smoke may be more effective and provide some steeringcapability.

4.4 Static Virtual Fence Experiments

Methodology In the final field experiment, we used a total of six collars to test the effects of the virtualfence on the herd. These collars were put on the cows with one virtual fence already present, allowing usto be sure that the fence would be present even if we experienced base-station communication failures. Thecows were sent into field 1, with a north-south oriented virtual fence located across the paddock about onethird of the way up. We observed the cows’ reactions visually to supplement the logged data. After thecows had moved through the preset fence and the fence had timed out, a second north-south fence wasinstantiated near the top of the paddock (approximately under the “1” label in Fig. 6). Both fences usedthe graduated volume algorithm, with a value of 7%/m for the first fence and 5%/m for the second.

Data Of the six collars, two performed very well for the duration of the experiment, two performed wellbut for a shorter time (perhaps due to battery failure) and two had poor to nonexistent GPS signal, probablydue to rotation of the collars on the cows’ necks. Figure 16 shows data from one cow’s collar over the entire

Cow Ori

enta

tion

Sound/Volu

me

Rea

ctio

ndir

ection

Rea

ctio

nm

agnitude

Comments10 6 air/50 6 1 step rapid 3 in a row, startled10 6 air/50 - - 2 in a row, nothing10 6 dog/40 - - lifted head, looked to one side10 6 dogx4/50 12 6 steps walked forward while sound was on3 - air/40 - - very startled, shuddered at each sound3 - air/26 - - no reaction3 - air/60 - - no reaction3 - air/80 - - no reaction, habituated ? (we were close

to the cow)3 - dog/50 - - no reaction8 12 airx6/50 12 walked started walking for duration of walk (ini-

tial 2s delay), actually moved toward us8 12 airx6/50 - - no reaction (her back to us)8 12 cymb/50 12 walked for duration8 12 dog/50 10 cow and neighbour moved8 12 dog/50 - - no reaction8 12 cymb/50 - - no reaction8 12 hiss/50 - - no reaction8 12 crash/50 - - no reaction, neighbours looked up8 12 air/100 - - flicked her tail8 12 dog/100 - - no reaction

Table 1: Observed reaction to stimulus. Time increases from top to bottom in the table. Orientation andreaction direction are specified using “hours on the clock” notation where noon is North. Volume is percentvolume setting on the Zaurus.

Figure 15: Velocity of cow 10 versus time with sound stimulus events noted by asterisks. There appears tobe good correlation between sound events and cow motion.

experiment. Both fences are shown here relative to the cow’s travels in the field. This figure shows thesounds being applied at the correct locations for both fences. We also show a closeup of another collar’sdata, showing that the fence worked correctly over multiple crossings, with the fence timeout resetting asdesired. In addition, to analyze the effects of the fence, we looked at the speeds of all the cows during thetimes sounds were being played relative to the rest of the time.

Discussion Our visual observations were that in general, the cows noticed the sounds, but either ignoredthem or did not make the desired association with their position. For two of the cows, we observed theanimal stop grazing when the sound was played, look up and walk slowly in a different direction. However,this new direction was not sufficiently different to take the cow into the interior area, and further soundsseemed to be ignored. We also observed one cow essentially ignore the sounds entirely. We were told thatthe cows tended to be motivated to reach the top of this paddock, especially first thing in the morning, andthis motivation may have been too strong for the sounds to overcome. However, the second fence nearer thetop of the hill was also not effective at keeping the animals on the desired side.

We also analyzed the logged data for the two cows that recorded good data for both fences. For bothcows, the logs seem to indicate that the first fence slowed the cows’ progress toward the top of the hill.This was determined by comparing for each cow (1) the cow’s speed between entering the field and reachingthe first fence and (2) the cow’s speed while the first fence was causing sounds to be played. For cow 10,the average speeds for these two time periods were 0.380 m/s and 0.255 m/s respectively, and for cow 9,0.590 m/s and 0.388 m/s respectively. For both, this difference is significant at the 0.01 level using a t-test,and the form of the speed distributions for these time periods looks quite similar. Later speed data is lessconvincing due to habituation to the stimulus. For cow 9, after the first fence stops making noise, up throughand including when the second fence makes noise, its speed did not change significantly, whereas for cow 10there was a speed increase between the fences and decrease for the second fence. For both cows, once theyhad reached the top of the field, their speed and range decreased significantly (again, using a t-test with a0.01 significance level), both just under 0.2 m/s on average, similar to the grazing speeds seen in the earlierexperiment.

0 100 200 300 400 500

50

100

150

200

250

300

350

400

450

500

East [m]

Nor

th [m

]

Cow motionLocations where sounds playedFences

Figure 16: Trace of the position of cow #10 during an experiment with two virtual fences. The cow startedin the barn, at the SE corner of the plot. Locations where sounds were played automatically are shown. Thelong straight line to the north is the walk from the barn to the field, which is not considered in our analysis.After the first fence (to the east) timed out, no sounds played until a second fence (to the west) was created.

−20 0 20 40 60 80 100 120360

380

400

420

440

460

480

East [m]

Nor

th [m

]

Figure 17: Trace of the position of cow #9 during the virtual fence experiment. A small portion of theexperiment is shown, during which the cow walked past the fence on two separate occasions, showing correctbehavior of the virtual fence algorithm.

−20 0 20 40 60 80

50

60

70

80

90

100

110

120

130

Motion along paddock corridor

East [m]

Nor

th [m

]

−20 −10 0 10 20−20

−15

−10

−5

0

5

10

15

Motion relative to virtual paddock

East [m]

Nor

th [m

]

(a) (b)

−50 0 50 10040

60

80

100

120

140

160

Motion along paddock corridor

East [m]

Nor

th [m

]

−50 0 50−30

−20

−10

0

10

20

30

40

Motion relative to virtual paddock

East [m]

Nor

th [m

]

(c) (d)

Figure 18: Dynamic fence experiment results. (a) Absolute and (b) relative motion of one person. (c-d)Absolute and relative motion data for four people.

4.5 Dynamic Virtual Fence Experiments

Methodology We have tested the dynamic fence algorithm using the smart collars on people. Plans werecreated in the planning GUI and exported directly to the collars. The experiment will be conducted withdifferent hardware and different animals at the USDA3 Jornada test range in New Mexico with the assistanceof Dr Dean Anderson, but those results are not yet available.

Data Motion traces for one person and the group of four people are shown in Fig. 18. Fig. 18(a) showsthe absolute coordinates of the paddock path and the person within the paddock. However if we plot theposition of the person relative to the instantaneous position of the paddock the situation is more interesting,as shown in Fig. 18(b). The schedule used to generate this experiment is showed below.

3US Department of Agriculture

(a) (b)

Figure 19: (a) The Fleck2 tracking board, and (b) incorporated into a collar.

fence-long fence-lat norm-x norm-y speed(m/s) t-start (s) t-end(s)-72.28808 43.70385 1.00 0.00 -0.1302 0 215-72.28808 43.70385 0.00 -1.00 0.2000 0 215-72.28788 43.70370 -1.00 0.00 0.1302 0 215-72.28788 43.70370 0.00 1.00 -0.2000 0 215-72.28843 43.70346 1.00 0.00 0.2000 215 400-72.28843 43.70346 0.00 -1.00 0.2000 215 400-72.28823 43.70331 -1.00 0.00 -0.2000 215 400-72.28823 43.70331 0.00 1.00 -0.2000 215 400

This schedule represents the fences used in the experiment portrayed in Fig. 18. The first four linesrepresent the four fences that form the initial safe square at the top of Fig. 18(a) and (c). Note that thenormal vectors for these four fences are along each of the four cardinal directions, and that they all startat time 0 and continue for the same amount of time. After 215 seconds, these fences reach the lines shownby the square in the middle of the plots. The second set of four fences exactly overlap the first four at thismoment, but have different velocities, so the safe square seamlessly moves off in a different direction, andthe first set of fences is no longer in existence to impede travel.

Discussion The results were successful algorithmically, in that the fences appeared in the correct real-worldlocations at the desired times, but were less successful from a herding standpoint. The absolute positiontrace shows that the general motion of the people was along the desired corridor, as in Fig. 18(a). Howeverfrom the paddocks perspective, the people spend very little time within its bounds, as seen in Fig. 18(b).This may be biased by the experimenters walking in the general direction of the virtual paddock motion.

One benefit of this experiment was that the subjects were able to verbalize their feelings about thesystem. Primarily, they felt that the resolution of the stimulus gradient was insufficient to be of assistance.Also, they tended to be inquisitive, actively exploring the acceptable boundaries of their space. Togetherwith relatively high walking speeds (averaging over 0.5 m/s) they were able to move a large distance betweensounds and were unable to relocate the virtual paddock. We believe that with a more appropriate stimulus(such as Anderson’s stimulus for cows [3]) that this can be overcome.

5 Conclusion

In this paper we have introduced the concept of a virtual fence, a device which applies a stimulus to an animalas a function of its pose with respect to one or more fencelines. The fence algorithm is implemented by a smallposition-aware computer device worn by the animal, which we refer to as a Smart Collar. We have describeda simulator based on potential fields and stateful animal models whose parameters are informed by field

observations and track data obtained from the Smart Collar. We have observed the effect of sound stimulion the animals but have had difficulties due to habituation, and the option of electric shock stimulus hasnot been available to us. Field implementations of static virtual fencelines have been tested and despite thelimitations of sound stimuli, a change in behaviour was observable. Dynamic virtual fences were implementedand tested on human subjects which gave some interesting perspectives which we cannot get from animalsubjects.

Our approach to modeling animal behavior is based on animal behavior literature, our limited empiricalobservations of cow behaviors, and our discussions with cattle farmers and experts from the USDA and fromHeytesbury Beef, Australia This simple model has proved to be effective. However, in the future we wouldlike to use the animal tracking data to establish more accurate behavior models and parameter.

Our future work has a number of different directions and different locations. We will work with researchersat the USDA in New Mexico to trial the dynamic virtual fence concept. In Australia we are trialling a newcollar [26] which is a variant of a Mote sensor network device, comprising an AVR microprocessor, 433MHzradio link, GPS, 3-axis acceleration, 3-axis magnetic field and temperature, programmable in TinyOS andwith the ability to hold considerable amounts of data in an MMC flash memory card, see Figure 19. Wealso wish to take animal position data and extract information about herd social structure and dynamics.This will permit more significant grounding of this work on real data which will lead to new and improvedmodels. These models will lead to a better understanding of cattle behavior and control at the individual andgroup level, which has the potential to impact not only the cattle industry, but more broadly, agriculture.By controlling grazing patterns, there will likely be an ecological impact as the land will be more carefullyutilized.

A Index to Multimedia Extensions

Extension Type Description1 Video Simulation of unforced animal motion2 Video Simulation of forced animal motion, slow threat object3 Video Simulation of forced animal motion, fast threat object4 Video Simulation of a virtual fence5 Video Simulation of two animal populations being separated6 Video Observed effect of stimulus on animals in the field

Acknowledgments

There are many people we would like to thank for assisting us in this work. Simon Holmes a Court inspiredus to think about this problem and and has given us great technical insights into cows and herding. StevePetty of Heytesbury Beef Australia and the staff at Pigeon Hole station showed great hospitality and pro-vided valuable discussion and suggestions. We especially thank Steven, Kerry and Paul at Cobb Hill fortheir assistance and enthusiasm and the use of their cows for these experiments. We thank Lorie Loeb forfacilitating our work at Cobb Hill Farms. We also received some valuable cow domain knowledge from Budand Eunice William’s Stockmanship School (stockmanship.com) who teach low-stress cow herding controlto farmers. Tom Temple assembled the cow collars. Jenni Groh provided invaluable advice on the exper-imental procedure. The experimental part of this work has been done under protocol assurance A3259-01given by the Institutional Animal Care and Use Committee (IACUC) of Dartmouth College. Dean Andersonof USDA Agricultural Research Service at the Jornada Test Range in New Mexico has been very helpful indiscussions regarding virtual fencing. Dave Swain of CSIRO Livestock Industry has been a keen supporterof the ongoing Australian cattle tracking experiments.

References

[1] Jean Marc Ame, Colette Rivault, and Jean Louis Deneubourg. Cockroach aggregation based on strain odourrecognition. Animal Behaviour, page in press, 2003.

[2] D.M. Anderson and C.S. Hale. Animal control system using global positioning and instrumental animal condi-tioning. Technical Report US Patent 6,232,880, USDA, May 2001.

[3] D.M Anderson, C.S. Hale, R. Libeau, and B. Nolen. Managing stocking density in real-time. In N. Allsop, A.R.Palmer, S.J. Milton, K.P. Kirkman, G.LH. Kerley, C.R. Hurt, and C.J. Brown, editors, Proc VII International

Rangelands Conf., pages 840–843, Durham, South Africa, August 2003.

[4] T. Balch. Hierarchic social entropy: an information theoretic measure of robot team diversity. Autonomous

Robots, July, 2000.

[5] T. Balch, Z. Khan, and M. Veloso. Automatically tracking and analyzing the behavior of live insect colonies. InAutonomous Agents, 2001.

[6] Tucker Balch and Ronald C. Arkin. Behavior-based formation control for multiple mobile robots. IEEE Trans-

actions on Robotics and Automation, 14(6):929–39, December 1998.

[7] Tucker Balch and Maria Hybinette. Social potentials for scalable multi-robot formations. In Proc. of Int’l Conf.

on Robotics and Automation, pages 73–80, San Francisco, April 2000.

[8] Z. Butler, P. Corke, R. Peterson, and D. Rus. Dynamic virtual fences for coltrolling cows. In in the 2004

International Symposium on Experimental Robotics, 2003.

[9] Z. Butler, P. Corke, R. Peterson, and D. Rus. Virtual fences for controlling cows. In ICRA, New Orleans, LA,2004.

[10] G. Caprari, A. Colot, R. Siegwart, J. Halloy, and J.-L. Deneubourg. Building mixed societies of animals androbots. IEEE Robotics and Automation Magazine, page in press, 2004.

[11] G. Caprari, A. Colot, R. Siegwart, J. Halloy, and J.-L. Deneubourg. Insbot: Design of an autonomous mini mobilerobot able to interact with cockroaches. In in the 2004 International Conference on Robotics and Automation,2004.

[12] Don Comis. The cyber cow whisperer and his virtual fence. Agricultural Research, November 2000.

[13] J. Gautrais, C. Jost, R. Jeanson, and G. Theraulaz. How individual interactions control aggregation patterns ingregarious arthropods. Interaction Studies, page in press, 2004.

[14] L. Howery, D. Bailey, and E. Laca. Impact of spatial memory on habitat use. Technical report, 1999.

[15] R. Jeanson, S. Blanco, R. Fournier, J.-L. Deneubourg, V. Fourcassie, and G. Theraulaz. A model of animalmovements in a bounded space. Journal of Theoretical Biology, 225:443–451, 2003.

[16] Philo Juang, Hidekazu Oki, Yong Wang, Margaret Martonosi, Li-Shiuan Peh, , and Daniel Rubenstein. Energy-efficient computing for wildlife tracking: Design tradeoffs and early experiences with zebranet. In Proceedings of

the Tenth International Conference on Architectural Support for Programming Languages and Operating Systems,2002.

[17] Alfred J. Lotka. Elements of Physical Biology. Williams and Wilkins Co., 1924.

[18] R.E. Marsh. Fenceless animal control system using gps location information. Technical Report US Patent5,868,100, Agritech Electronics, February 1999.

[19] A. Martinoli, G. Theraulaz, and J.-L. Deneubourg. Quand les robots imitent la nature. La Recherche,358(november):56–62, 2002.

[20] Maja J. Mataric. Kin recognition, similarity, and group behavior. In Proc. of the Fifteenth Annual Cognitive

Science Society Conference, Boulder, Colorado, 1993.

[21] P. Michelena, P. M. Bouquet, A. Dissac, V. Fourcassie, J. Lauga, J.-F. Gerard, and R. Bon. Do female and malemerino sheep segregate within the same ecological conditions. Animal Behaviour, page in press, 2003.

[22] Lynne E. Parker. ALLIANCE: An architecture for fault tolerant, cooperative control of heterogeneous mobilerobots. In Proc. of IEEE Int’l Conf. on Intelligent Robots and Systems, pages 776–83, Munich, Sept. 1994.

[23] T.M. Quigley, H.R. Sanderson, A.R. Tiedemann, and M.K. McInnis. Livestock control with electrical and audiostimulation. Rangelands, June 1990.

[24] John H. Reif and Hongyan Wang. Social potential fields: A distributed behavioral control for autonomousrobots. Robotics and Autonomous Systems, 27(3):171–94, May 1999.

[25] S. Russell and P. Norvig. Artificial Intelligence: A Modern Approach. Prentics Hall series on Artificial Intelligene,2002.

[26] Pavan Sikka, Peter Corke, and Leslie Overs. Wireless sensor device for animal tracking and control. In First

IEEE Workshop on Embedded Networked Sensors within the 29th Conference on Local COmputer Networks,pages 446–454, 2004.

[27] B. Smith. Moving ’Em: A Guide to Low Stress Animal Handling. Graziers Hui, 1998.

[28] A.R. Tiedemann, T.M. Quigley, L.D. White, W.S. Lauritzen, J.W. Thomas, and M.K. McInnis. Electronic(fenceless) control of livestock. Technical Report PNW-RP-510, United States Department of Agriculture,Forest Service, January 1999.

[29] R. Vaughan, N. Sumpter, A. Frost, and S. Cameron. Robot sheepdog project achieves automatic flock control.Proc. Fifth International Conference on the Simulation of Adaptive Behaviour., 1998.

[30] I. Veissier, A. Booissy, R. Nowak, P. Orgeur, and P. Poindron. Ontogeny of social awareness in domesticherbivores. Appl. Anim. Behav. Sci, (57), 1998.